Micro-topography driven vegetation patterns in open mosaic landscapes

Abstract

Elevation models based on remotely sensed data, especially high-resolution Digital Terrain Models (DTMs) generated using airborne laser scanner (ALS) data, are increasingly being used for the analysis of plant diversity patterns in open landscapes. The vegetation pattern of alkali landscapes shows a high correlation with the position of water table and salt accumulation, which are strongly correlated with topographic variations occurring at a small spatial scale of a few decimetres (micro-topography). In this study we classified eight grassland associations in an alkali landscape based on a DTM generated from ALS data at a pixel size of 0.25m, and 30 variables derived from the DTM, using an ensemble learning method (Random Forest). Our aim was to identify the micro-topographic variables which could be indicators of vegetation pattern in alkali landscapes. The associations range from Cynodon pastures (short dry grasslands on soil with low salt content) occupying the highest elevations to Beckmannia meadows (wet grasslands on soils with moderate salt content composed of tall grass species) at the lowest elevations, with an elevation difference of approximately 1.2m between the two. Apart from slope, aspect and curvature, we used Topographic Wetness Index (TWI), and Topographic Position Indices (TPI) at various kernel sizes ranging from 50cm to 500m for the classification. The eight associations were also grouped into four aggregated categories — loess grasslands, alkali steppes, open alkali swards and alkali meadows — for further analysis. Vegetation of the studied alkali landscape could be classified into the eight associations with an accuracy of κ: 0.56, and into the four aggregated categories with an accuracy of κ: 0.77 using all the variables. Sequential backward and forward selections of variables were implemented to reduce the number of variables while maximising the accuracies, resulting in increased accuracies of κ: 0.72 and κ: 0.83 for the associations and aggregated categories using six and three variables respectively. TPI at different kernel sizes, previously used to explain vegetation distribution in mountainous areas, was found to be a better indicator of vegetation types than absolute elevations in lowlands where the elevation differences are more subtle. Two characteristic features of the study area — erosional channels and alkali steps — could also be delineated using micro-topographic variables. The results point to the possibility of large-area mapping and monitoring of grasslands where micro-topography is an indicator of vegetation, using only the elevation data from ALS.